The Rise of AI Companions: Interaction with AI Companions and Psychological Well-being
Yutong Zhang, Dora Zhao, Jeffrey T. Hancock, Robert Kraut, Diyi Yang
cs.HC
2025-06-15
Across 1,131 Character.AI users and 464,687 messages, overall chat intensity linked to higher well-being, but companionship-oriented use tracked with lower scores, sharpest among high self-disclosers.
Companion chatbots like Character.AI have grown expressive enough that users form relationship-like bonds with them. The unresolved question is whether these bonds ease loneliness or make things worse. Existing studies are contradictory and lean almost entirely on surveys, with no view into what people actually say to their bots. This paper stitches together survey responses, free-text relationship descriptions, and donated chat logs to map how different ways of using these bots line up with well-being.
The team recruited 1,131 U.S. adult Character.AI users through Prolific. Of these, 237 donated their chat histories: 4,664 sessions and 464,687 messages. They measured companionship use three ways, a forced-choice primary-use question, free-text descriptions of the bot relationship, and the share of companionship-type sessions in the chat logs, classifying with GPT-4o, summarizing with Llama 3-70B, and theme-mining with TopicGPT. Interaction intensity came from an adapted Facebook Intensity Scale, self-disclosure from a MOCA subscale, and offline network size from the LSNS. Well-being used six items from the Comprehensive Inventory of Thriving (life satisfaction, positive and negative affect, loneliness, social support, belonging). Associations were estimated with standardized-coefficient linear regressions and 95% confidence intervals.
The counterintuitive contrast: more overall chatting tracked with higher well-being (β=0.27, 0.29, p<.001), yet companionship-oriented use tracked with lower well-being, consistently across all three measures.
| Measure | Link to well-being |
| Overall intensity | β=0.27 (higher) |
| Companionship (primary use) | β=−0.48 |
| Companionship (relationship text) | β=−0.32 |
| Companionship (chat share) | β=−0.27 |
The negative link was stronger among intensive users (β=−0.31) and high self-disclosers (β=−0.38). Smaller offline networks predicted companionship use (β=−0.03).
Prevalence is itself telling: only 11.8% picked companionship as their primary use, yet 51% used words like "friend" or "partner" in free text, and 92.9% of donated logs contained at least one companionship session. By theme, emotional and social support appeared in 80.3% of sessions, romantic and intimacy roleplay in 68.0%, and dark or risky content in 30.7%.
For anyone building companion or social AI, the takeaway is that the effect depends on how the product is used, not how much. The split is between treating the bot as a tool versus a relationship substitute, sharpened by how deeply users disclose. The authors argue these systems should be framed as limited tools that build social skill, not as connection replacements. For trust-and-safety teams, the 18% of sessions surfacing self-harm-related content is a concrete guardrail signal.
The authors' own headline caveat: this is cross-sectional data, so there is no causal claim. Companionship use may erode well-being, or lower-well-being users may simply gravitate to it. The data leans on self-report, chat donation was voluntary and selective, and the sample skews toward digitally literate users on a single U.S. English platform. Character.AI also changed over the Nov 2022 to Jan 2025 window, and the authors had no visibility into moderation-rule shifts.
One more discount from reading it closely: the largest coefficient (β=−0.48) comes from the forced-choice measure that only 11.8% selected, and several models have R² near 0.03. The network-size link (β=−0.03) is significant but tiny in effect.